Our expected mode of action depends on whether we feel relaxed or depressed, excited or anxious. For social animals that care for each other and work together – like us – it is crucial to track each other's feelings. A mother responds to the distress of her baby; a child checks whether its parents approve; a coach gives a pep talk when team morale is low.

In this line of work, we use active inference – a next-generation, biology-based form of (Bayesian) artificial intelligence – to model and understand how (shared) affect and more complex emotions can emerge in hierarchical living systems like ourselves. Under this approach, deep biological systems form emergent affective beliefs based on evidence gathered from lower levels (“bottom-up”) and reduces uncertainty about such inferences through internal modulation of expectations (“top-down”). We discuss the utility of such models as blueprints for – among other things – computational neuroanatomy, precision psychiatry and life-like artificial intelligence.